We combine these models to predict average total response time R for the service, deliberately ignoring congestion on the network paths to the clients (which the utility OS cannot control). Given a measure DP of average per-request service demand on the Web server CPU, CPU response time RP is given by a simple queuing model similar to the storage model above; previous work [12] illustrates use of such a model to adaptively provision CPU resources for Web services. The service's average response time R is simply:
R = RP + RS (1 - H) | (5) |